Highlights • An efficient method is proposed for accurate traffic demand forecasting. • It has three main stages: deep learning, error analysis, and optimal reconciliation. • It relies on a hierarchical time series (HTS) structure to find coherent forecasts. • The deep-learning method is improved for area-based demand prediction via HTS. • The error analysis is vital to do the reconciliation in the HTS structure optimally.

Abstract Mobility services require accurate demand prediction in both space and time to effectively manage fleet rebalancing, provide quick on-demand responses, and enable advanced ride-sharing with minimum fleet size. Although the optimization of mobility services is a widely studied topic, the demand prediction side has received relatively less attention. In this paper, we aim to develop an efficient method for traffic demand forecasting by means of deep learning and hierarchical reconciliation approaches. The concepts, as well as the theories behind the proposed approach, are founded on Hierarchical Time Series (HTS), which also adopts Long Short-term Memory (LSTM) as a special kind of Recurrent Neural Network (RNN) to learn from the associated time series and produce reliable demand predictions. The proposed approach relies on the proper design of the HTS structure to get accurate forecasts for the number of trip departures and their associated uncertainty within predefined zones and aggregated collections of these zones. Moreover, the error analysis is essential for accomplishing the reconciliation in the HTS structure optimally. The proposed approach, consisting of deep learning, error analysis, and optimal reconciliation stages, has a remarkable ability to predict demand and control the forecasting quality at all levels of the hierarchical structure. We evaluate the proposed approach using a large-scale GPS tracking dataset from Lyon, France. The proposed method reduces the root mean square error (RMSE) by 13.92% and 14.77% for predefined and aggregated zones, respectively, compared with the LSTM using the historical demand and external features of time at the quarter-hour time step. Similarly, the corresponding improvements in mean absolute percentage error (MAPE) are 14.87% and 19.23%, respectively.